Solutions / Catalogue / Acoustic & Audio AI

Acoustic & Audio AIReference design

Retail Voice

A speech-to-text system for shop counters, meeting rooms and consultation rooms that stores conversations as searchable text with names and numbers masked. Audio comes from a table microphone array or a wearable Clip, and recognition runs on a local box or on a store server. It is aimed at chain stores, clinics, insurance offices and training teams that want records without keeping raw audio for long.

RetailHealthcare

Talk to an engineer See the hardware

Indian jewellery showroom with glass display counters and gold necklaces on shelves, marigold garlands overhead — Photo: JL Gold Palace, CC BY-SA 4.0, Wikimedia Commons. Shows the setting, not this installation.

4.82%: Chinese character error rate on 100 AISHELL-1 public dataset clips
0.98 / 0.95: Precision and recall for masking names and numbers on a 114-sample Chinese and English test set
167 ms: Median time from end of speech to final text on J3011 with 32 streams (vendor test)
505 ms: Median end of speech to text on J4012 with 48 streams (vendor test)

Results reported by the design's authors on their own test setups. Yours depend on the site, so we confirm them in a pilot.

The problem

Why it matters.

Conversations at counters, meetings and consultations are not captured as searchable text.

The flow

From signal to action.

01SenseAudio captureA reSpeaker XVF3800 array on the counter, or a reSpeaker Clip via a phone app, picks up speech.
02DecideRecognitionSpeech is turned into text on a local host, or on a server after the audio is uploaded.
03DecideMaskingNames and numbers are swapped for placeholders before anything is saved.
04RecordStorageText goes to MySQL, MinIO and transcript files, and raw audio is kept for a short window.
05PeopleConsoleStaff search, export and erase records in a browser console or through APIs.

Deployment architecture

How it is wired.

The devices on site, what carries their data, where the decision is made and where the result goes.

01SenseOn site

reSpeaker XVF3800 USB 4-mic array1 per counter or room
reSpeaker Clip1 per wearer

02ConnectCarries the data

reRouter CM4 (4 GB RAM, 32 GB eMMC)1 per site
Bluetooth LE

03DecideRuns the model

reComputer RK3576-301 per site
reComputer J4012
reComputer J3011
reComputer RK3588
Whisper
SenseCraft Voice App

04Your systemsWhere results land

WebSocket
HTTP API
MySQL
MinIO
Docker

How it works

Step by step.

01Choose where text should go and which audio source to use, then pick a hardware option by cost.
02Place one mic array per counter or room, within about 3 m of the speakers.
03Deploy the on-device or server preset and let models download on first start.
04Speech is transcribed, masked and stored as text, with uploads queued if the link drops.
05Check by speaking a sentence and reading the result, then set a retention period.

The hardware

What goes on site.

Items we stock can go straight in the bag. Everything else we source and quote with the installation.

reRouter CM4 (4 GB RAM, 32 GB eMMC)Lowest-cost pilot transcriber on CPU1 per siteSourced on quote
reSpeaker XVF3800 USB 4-mic arrayMicrophone array for one talking area1 per counter or roomSourced on quote
reComputer RK3576-30NPU transcriber or server stack host1 per siteSourced on quote
reSpeaker ClipWearable recorder that uploads via a phone app1 per wearerSourced on quote
reComputer J4012Higher-capacity GPU transcription hostSourced on quote
reComputer J3011GPU transcription host with the lowest measured latencySourced on quote
reComputer RK3588Alternative NPU transcription hostSourced on quote

Before you install

Speakers within about 3 m of the mic array.
Steady background noise at or below 70 dB.
One array and host per talking area. A second counter needs its own set.
A server host with an arm64 CPU and 20 GB of free disk.
Store LAN plus an uplink to the backend, and internet on first start to fetch models.

Good to know

Accuracy drops beyond 3 m or above 70 dB, and better hardware will not fix that.
Raw audio is kept for 24 hours by default and is not masked.
Spoken-out numbers are only caught for Chinese mobile numbers, so Indian phone numbers said as words may not be masked.
Both presets are marked Beta, and MinIO's AGPL licence should be checked for commercial use.

How we deliver it

Paper to site, one team.

01Site surveyWe walk the site, or review photos and drawings, and agree what to measure and where each device goes.
02PilotOne lane, one line or one room first. We tune it on your conditions and agree what good looks like.
03Roll outWe supply the hardware in India with GST invoices, install it, and connect it to the systems you already run.
04SupportSpares, remote checks and model updates, with one team to call when something changes.

Questions

Asked before buying.

Does it understand Indian languages?

The page only reports tests in Chinese and English. Hindi and regional languages would need testing first.

Can we delete a customer's records on request?

Yes, the server stack has an erase API. The on-device preset has no delete API, so its folder must be cleared by hand.

How noisy can the store be?

Up to about 70 dB of steady noise, with speakers within 3 m of the array.

Want this on your site?

Tell us the site and the problem. We come back with a pilot plan and a priced hardware list.

Book a call with an engineer Write to us

Want this on your site?

In-Person Conversation Intelligence

Counter Audio Capture & Analytics

AI Voice Warehouse Management

Want this on your site?

In-Person Conversation Intelligence

Counter Audio Capture & Analytics

AI Voice Warehouse Management